Data science is a major area of investment for banks due to its proven impact on operations such as fraud protection, risk mitigation, customer relationship management, and more. But while investments in AI are growing, banks are often finding that their existing analytics and business intelligence technology and talent aren’t capable of meeting their current and expanding needs. Challenges in resources, technology infrastructure, and the ability to operationalize models quickly and efficiently can prevent financial institutions from fully leveraging AI and data science to drive business impact.
These challenges, paired with the need to remain competitive in a quickly evolving market, compelled the Sumitomo Mitsui Banking Corporation (SMBC) to seek out innovative solutions to help it maximize its AI and machine learning (ML) investments.
ADVANCED THINKING, LIMITED BY LACK OF RESOURCES : SMBC was formed in 2001 by the merger of Sumitomo Bank and Sakura Bank. With over $1,775.14B in assets, SMBC is the world’s 14th largest bank and provides offerings across a broad spectrum of financial services including consumer banking, corporate and investment banking, international banking, and more.
In early 2016, SMBC’s IT Planning department was tasked with addressing a growing concern for SMBC: while the bank had begun using machine learning in several of their business divisions — for operations, such as enhancing customer product upsell and cross-sell opportunities, managing customer attrition and identifying default risk, SMBC’s nascent data science team was facing a shortage of talent faced with an excess of demands.
While building ML and AI models was feasible, it was a 100% manual operation that required a lot of coding and data manipulation. Building a single model typically took two to three months with as much as 80% of the time spent on the process of creating the complex multi-dimensional flat tables required by ML and AI models. This process, known as feature engineering, coupled with the complexity and time-consuming nature of ML model selection and optimization, was hampering the ability ofSMBC’s team to deliver on all the projects required of them.
The combination of not enough talent, the complexity of models, and the time-consuming nature of feature engineering prevented SMBC’s team from scaling their data science practice, and restricted its output to only five new ML models in any given calendar year, with bandwidth to update an additional five models in that same timeframe.
AUTOML 2.0: FULL CYCLE DATA SCIENCE AUTOMATION AT SMBC : SMBC’s team, led by Akinobu Funayama and Tomohiro Oka, decided that AutoML technology was a possible solution to the broad lack of talent, and one that might help them accelerate development lifecycles for their data science projects. Critical goals for SMBC were the ability to analyze and optimize business models quickly and automatically, as well as the ability to automate as much of the data science lifecycle as possible. During the evaluation process, SMBC identified the automated creation of features, also known as “automated feature engineering” as a critical requirement for their project. Automating this manual and time-consuming process would enable SMBC to optimize resources and shorten project timelines. An additional crucial element for the use of these AutoML platforms was the need for transparency, to enable SMBC’s data science team to provide a higher level of transparency to business units that were asking for ML and AI applications.
To identify the best possible technology providers, SMBC explored more than 300 platforms, ultimately short-listing 50 providers to evaluate in more detail. SMBC ranked the short-listed platforms by their ability to meet the essential requirements of automated feature engineering and AutoML as well as ease of use for less experienced users. Several of the short-listed vendors were then tested in purpose-specific proof of concepts.
After their extensive evaluation cycles, SMBC decided that a combination of two separate types of platforms, used together, would best fit their needs.The first was a smart interactive data preparation platform that would help cleanse the master data. The second one was a data science automation (AutoML 2.0) platform that would automate the full-cycle of feature engineering, the selection, and optimization of ML models and also provided vital capabilities to help explain features associated with ML models to non-technical users.
48X ACCELERATION TO CHANGE THE GAME : Once implemented, the benefits achieved as a result of investing in AutoML 2.0 were immediate and substantial. Prior to implementing the technology, it took two months for data scientists to explore 2,000 features for each project during their development process. Through data science automation, SMBC can now examine more than two million features for each project. The benefits of automating both the feature engineering as well as the machine learning have also allowed SMBC to reduce development times dramatically. SMBC has cut its data science development times from an average of two to three months per project to less than 10 hours per project — a 48X acceleration in development times. The ability to explore the huge number of feature hypotheses has also improved the accuracy of models by as much as 30%, providing additional data insights which augment SMBC’s core domain knowledge.
The adoption of data science automation and the associated acceleration of development times has provided significant benefits to multiple business units at SMBC. Data scientists at SMBC no longer struggle with managing data and creating features, and can instead focus on feature analysis and on identifying models that are likely to be most beneficial to their business units. SMBC has also increased the number of models built per year to 100 models from 10 — a growth of 10X over the previous manual process. Finally, more transparency in feature engineering is allowing the team to explain ML models more succinctly to business units and to accelerate the feedback loop to continue to improve models over time.
SMBC: USING DATA SCIENCE AS A COMPETITIVE EDGE : As a result of their investment in data science and AutoML 2.0, SMBC has been able to expand the number of supported use-cases for data science and ML applications across the organization. In addition to customer management and marketing and sales, SMBC now uses AI and data science in numerous other divisions within the institution, including compliance, risk management, and financing and loans. This expansion has happened without having to add additional headcount to the data science team and has allowed SMBC to broaden the use of data science while also improving the accuracy of their models.
SMBC’s adoption of AutoML 2.0 has enabled them to automate many aspects of its data science process that were formerly time and resource-consuming. As a result, SMBC has been able to rapidly scale their AI/ML initiatives to drive transformative business changes.
originally posted on Forbes.com by Ryohei Fujimaki
